Suggestion Service

The suggestion service should not invent answers or imply automatic compliance. It should retrieve relevant approved datapoints and source material, draft a response, and show why the response is supportable.

Expected Behavior

• Search uploaded documents and structured datapoints together.
• Prioritize user-confirmed or reviewer-approved datapoints over raw extraction.
• Cite the exact source evidence and framework mapping used.
• Show when evidence is missing, stale, conflicting, or not mapped to the target standard.
• Distinguish CDP questionnaire support from external regulatory or standards compliance.
• Require human review before formal disclosure submission.

Retrieval Pattern

The useful pattern is not “ask an LLM to answer a disclosure question from scratch.” It is:

1. Resolve the discloser through Discloser Profiles.
2. Retrieve uploaded evidence and prior approved source material for that profile.
3. Retrieve structured datapoints attached to that profile.
4. Filter or rank datapoints by CDP question, reporting period, review state, and framework tags.
5. Draft a suggested response with citations and caveats.
6. Let the user review, edit, accept, or ignore the suggestion.

Relationship To Briink Disclosure Automation

The Briink disclosure automation project covers the uploaded-document and AI-assisted questionnaire pre-filling pattern. This Discloser Profiles opportunity is adjacent but distinct: it describes the internal CDP operating layer that would make extracted evidence reusable across profiles, standards, and future workflows.

In practice, Briink or another suggestion-service partner could extract candidate answers from documents, while Discloser Profiles and a datapoint layer could help CDP preserve identity, source, framework mapping, review state, and reuse boundaries.

Guardrails

• Do not present a mapped datapoint as proof of regulatory compliance.
• Do not hide source evidence or review status from the user.
• Do not reuse values across related entities unless the reporting boundary supports it.
• Do not prefer AI-extracted text over reviewer-approved structured datapoints.
• Do not treat framework mappings as product-ready unless the owning standards team maintains them as live metadata.